RESEARCH PAPER
Segmentation based building detection approach
from LiDAR point cloud
Anandakumar M. Ramiya
a,*, Rama Rao Nidamanuri
a, Ramakrishan Krishnan
ba
Department of Earth and Space Sciences, Indian Institute of Space Science and Technology, Department of Space, Thiruvananthapuram, Kerala, India
b
Amrita Vishwa Vidyapeetham, Coimbatore, India
Received 19 January 2016; revised 1 April 2016; accepted 4 April 2016
KEYWORDS LiDAR; Remote sensing; Building detection; Segmentation; PCL
Abstract Accurate building detection and reconstruction is an important challenge posed to the remote sensing community dealing with LiDAR point cloud. The inherent geometric nature of LiDAR point cloud provides a new dimension to the remote sensing data which can be used to pro-duce accurate 3D building models at relatively less time compared to traditional photogrammetry based 3D reconstruction methods. 3D segmentation is a key step to bring out the implicit geomet-rical information from the LiDAR point cloud. This research proposes to use open source point cloud library (PCL) for 3D segmentation of LiDAR point cloud and presents a novel histogram based methodology to separate the building clusters from the non building clusters. The proposed methodology has been applied on two different airborne LiDAR datasets acquired over part of urban region around Niagara Falls, Canada and southern Washington, USA. An overall building detection accuracy of 100% and 82% respectively is achieved for the two datasets. The performance of proposed methodology has been compared with the commercially available Terrasolid software. The results show that the buildings detected using open source point cloud library produce compa-rable results with the buildings detected using commercial software (buildings detection accuracy: 86.3% and 89.2% respectively for the two datasets).
Ó2016 National Authority for Remote Sensing and Space Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/ by-nc-nd/4.0/).
1. Introduction
Realistic 3D city models are gaining prominence with the dynamic growth of urban landscape (Tner, 1999). Various applications in facility management, utility management,
disaster management, noise modelling, and city planning (Biljecki et al., 2015; Sun and Salvaggio, 2013; Dadras et al., 2015) substantially benefit from the availability of 3D city models. There is an increase in the demand for street level 3D view of the urban landscape due to the commercial exploitation of 3D geospatial data based products by corpo-rate such as Google and Apple (Anguelov et al., 2010). The traditional photogrammetry based 3D capturing technique, which requires stereo images and extensive processing to create a 3D surface, is being replaced by the rapidly growing LiDAR
* Corresponding author.
E-mail address:[email protected](A.M. Ramiya).
Peer review under responsibility of National Authority for Remote Sensing and Space Sciences.
H O S T E D BY
National Authority for Remote Sensing and Space Sciences
The Egyptian Journal of Remote Sensing and Space
Sciences
www.elsevier.com/locate/ejrs
www.sciencedirect.com
remote sensing techniques for fast capture of the surface geom-etry (Baltsavias, 1999; Burtch, 2002). The LiDAR remote sens-ing is a fast and relatively cost effective means to capture and represent the realistic three dimensional structure of surface objects. LiDAR remote sensing systems record position and elevation of the target points asx,y,zcoordinates for represen-tation and storing (Baltsavias, 1999). The LiDAR points are recorded in an unorganized fashion due to the nature of the scan, requiring specific data organization and processing tech-niques which are significantly different from the existing image processing algorithms. The airborne LiDAR data captured over an urban setting consists of returns from both the natural (trees, bare earth) and man-made features (buildings, cars, roads, etc.). Building features identification and extraction is a key step in urban modelling.
Over the past decade, several algorithms for extraction of buildings from point cloud data have been reported by researchers (Zhang and Lin, 2012; Lari et al., 2011; Wang and Tseng, 2011a; Lafarge and Malllet, 2012). Building extrac-tion from the point cloud involves two sequential steps, namely, filtering, and segmentation. Filtering separates the ground points from non-ground points. There are many well established filtering algorithms available such as morphologi-cal filter (Vosselman, 2000), progressive densification filter (Axellson, 2000), surface based filter (Pfeifer, 2005), and seg-mentation based filter (Filin and Pfeifer, 2006). Morphological filter is based on using structural element which describes the admissible height difference between a ground point and a neighbouring point as a morphological operator to separate the ground and the non-ground points. In case of progressive densification, seed points are chosen to represent the ground points and are triangulated. The ground point set is progres-sively densified by finding the offset distance and angle of each of the points to the triangulated surface. In contrast to the above filter, surface based filter initially assumes that all the points belong to the ground and then removes non-ground points based on the weights assigned to the points. Segmenta-tion based filter first groups the points into segments using the local normal of each point, computed using the local neigh-bourhood. Segments are also created using region growing techniques. The segments are then classified into ground and non-ground based on their properties. A review of filtering algorithm byMeng et al. (2010)suggests that most of the filter-ing algorithms perform well for flat surfaces while producfilter-ing unacceptable results for undulating terrain.
Segmentation is applied on the non-ground points to detect the various objects present in a scene. Several 3D segmentation algorithms are available to detect landscape features from 3D point cloud. Density based segmentation is based on defining a neighbourhood of radiusrand all the points within the sphere of radiusrare said to belong to one cluster. Euclidean clustering based segmentation is one of the examples of density based seg-mentation (Rusu, 2009; Ghosh and Lohani, 2011). Region growing segmentation is based on growing a set of seed points based on the criteria such as global planarity and surface smoothness (Rabbani et al., 2006). Colour based region grow-ing method uses spectral information in addition to the geomet-rical information for segmentation (Ramiya et al., 2016). Incremental segmentation based on octree structured voxel space is based on establishing neighbourhood of the point cloud followed by coplanar segmentation and co-surface grouping (Wang and Tseng, 2011b). Adaptive segmentation (Lari et al.,
2011) is based on defining an adaptive cylinder depending on the point density followed by plane detection. Amongst these algorithms only the density based algorithm can detect both manmade and natural objects in a scene whereas the other algo-rithms are suitable only for manmade objects (planar objects). For building detection purpose, from the segments created, it is necessary to separate building segments from the non building segments. Limited literature is available on the methods of auto-matically separating a tree cluster from a building cluster. The existing studies are based on evaluating the fitness of the roof surface points on a plane. The clusters which satisfy these con-ditions are classified as building segments.
Apart from the key algorithms necessary for processing and analysis, implementation platform such as software platform is a key resource for the effective and affordable analysis of LiDAR point cloud. Commercial LiDAR processing software such as Terrasolid offers routines to automatically detect the building points from the input LiDAR point cloud. However, there is no defined method for automatically extracting build-ings from a point cloud besides the lower affordability of com-mercial software for many researchers. Recently many open source libraries are available to process 3D data. However there is no literature available showcasing the utility of the open source library for building detection. The objective of this work is to develop and validate a building extraction methodology using open source software tool. In this research, an open source point cloud library (PCL) has been used to cluster the entire point cloud data into segments. A novel building detection algo-rithm is then employed to separate the building clusters from the non building clusters. The buildings extracted using the pro-posed method are compared with the buildings extracted from a popular commercially LiDAR processing software.
2. Methods and materials 2.1. Filtering
Filtering is an important pre-processing step which separates ground points from the non-ground points thereby reducing the data size and helping in identifying building points. The adopted filtering algorithm belongs to the category of surface based filtering. Each point was given a weight based upon the distance to the mean interpolated surface. A threshold was determined based on the distance of each point from the mean surface. Based on the threshold value, the points are classified either as ground or non-ground point.
2.2. Data structuring
Each point of the LiDAR point cloud is organized in a file in the same pattern of the sensor scan mechanism. In general, the points in LiDAR data are discrete by space and unorganized due to the nature of geometry of the scanning device and target interaction. Hence the LiDAR data do not fit to represent on a grid and have an ill defined boundary hence called as point cloud. Due to this, all the three coordinate values (X,Y,Z) are required for encoding each of the point data. The random point cloud hence becomes more difficult to work with when performing operations such as search or for performing inter-polation. We organized the point cloud in a hierarchical data structuring using method kd tree method (Rusu, 2009).
2.3. Segmentation of the point cloud using PCL
3D segmentation of point cloud enables partitioning spatially isolated regions. We implemented the Euclidean distance based segmentation algorithm using PCL (point cloud library) which is an open source VC++ library (Rusu and Cousins, 2011). This segmentation method creates clusters based upon the Euclidean distance between a point and every other point. If the distance is within a particular threshold, the point is placed in a new cluster; otherwise the point is placed in a queue. The process continues until all the points are processed and no point remains in a queue. Minimum and maximum number of points in a cluster is user defined and hence the number of clusters varies for a particular dataset. As mentioned in the previous section, Euclidean distance based segmentation can detect both natural and manmade objects which makes it applicable across many types of landscapes.
2.4. Building cluster detection
To recreate buildings in the scene, the building clusters have to be separated from the clusters of other natural features such as trees. One simple method is to open each of the clusters sepa-rately in LiDAR visualization software and separate the build-ing and the non-buildbuild-ing cluster. However, this process is cumbersome and time consuming. Automation of this process led to a novel building detection algorithm based on the his-togram of the local normal of each point in the clusters. In this method, the local surface normal was computed for each of the points in the cluster. The direction cosines of the normal were found out and histogram was generated. The statistical param-eters of the histogram such as mean, range and standard devi-ation were computed. These parameters vary significantly for a tree cluster from that of a building cluster. This methodology can be adopted to separate building cluster from non-building cluster.
2.5. Building detection using commercial software
Terrasolid is a commercial software capable of end to end pro-cessing of LiDAR point cloud. This software uses the progres-sive densification algorithm for filtering, i.e., separating ground points from non-ground points. The ground points which are identified from the filtering process are used to clas-sify the rest of the points. The non-ground points are labelled into various land cover categories such as low vegetation, high vegetation, buildings based on the distance of the point with respect to the ground points. To detect the set of building points from the non-ground points, the algorithm starts by identifying the holes in the ground points. The non-ground points which are in the location of the holes are checked for planarity conditions. If it satisfies, the points are classified as building points (Soininen, 2015).
3. Experiment on LiDAR datasets
The proposed methodology was implemented using the open source point cloud library in Visual C++ on two different air-borne LiDAR datasets which differ by point density and return numbers. The first dataset (dataset 1 in Table 1) was
collected in 2004 over the Niagara Falls’ neighbourhood using the Airborne Laser Terrain Mapping (ALTM) 3100 sensor (Optech Inc., Concord, Canada) at a flying height of 1190 m.
Table 1 Airborne LiDAR datasets used for the study.
Parameters/dataset Dataset 1 Dataset 2 Pt density (pts/sq m) 2 5.5 Number of 1 return 31,033 108,198 Number of 2 return 2594 0 Number of 3 return 76 0 Number of 4 return 0 0 Total number of points 33,703 108,198
xrange 112 103
yrange 139 201
Figure 1 Before and after filtering: (a) dataset 1 before filtering, (b) non-ground points from dataset 1.
Figure 2 Before and after filtering: (a) dataset 2 before filtering, (b) non-ground points from dataset 2.
The density of the data is 2 points/sq m. The second dataset (dataset 2 inTable 1) was part of the LiDAR data collected in 2005 over the Yakima county of southern Washington using the Terrapoint-s40 ALTMS flying at a height of 1060 m. The density of the data is 5.5 points/sq m.
4. Results and discussion
Figs. 1 and 2show the output after filtering of both LiDAR datasets. The ground points are effectively removed from the point cloud (seeFigs. 1 and 2).Table 2shows the details of points before and after filtering. The filtered dataset was arranged using kd tree structuring. The points were then seg-mented into different clusters by 3D Euclidean distance based segmentation. From the clusters obtained, building and non building clusters were separated using the histogram of the local surface normal of each cluster.
Each of the clusters obtained for the dataset was catego-rized into building and non-building cluster based on the peaks of the respective histogram.Fig. 3shows the histogram of the
non-building cluster. The range and mean of the values of the elements in the histogram can be used to separate a building cluster from a non-building cluster. As seen inFig. 3, the his-tograms derived from the building clusters exhibit few distinct peaks whereas the histogram of the non-building clusters spread wide apart. For dataset 1, 35 clusters were identified. Some of the user defined parameters included minimum and maximum number of points in the clusters. In this case, the minimum cluster size was 50 and maximum cluster size was 5000. Visual inspection of the dataset 1 indicates that there are 25 buildings and 3 trees. The method implemented in this work identified 25 buildings as can be seen inFig. 4.
Table 2 Number of points before and after filtering.
Parameters LiDAR dataset
Dataset 1 Dataset 2 Total number of points 33,703 108,198 Ground point 22,645 72,478 Non-ground point 11,058 35,720
Figure 3 Histogram of the surface normal of a building cluster and a non-building cluster (a) histogram of surface normal of tree in thex
direction, (b) histogram of surface normal of tree in the y direction, (c) histogram of surface normal of tree in thezdirection, (d) histogram of surface normal of building in the x direction; (e) histogram of surface normal of building in theydirection, (f) histogram of surface normal of building in the z direction.
Figure 4 Result of Euclidean based segmentation (a) top view of building clusters for dataset 1 (b) building clusters for dataset 1.
Although all the buildings are identified by the segmenta-tion process, some buildings were segmented as multiple clus-ters. The mark ‘MBC’ represents the multiple building clusclus-ters. One reason which can be attributed to this is the presence of multiple roofs for the building at different heights. These par-tial buildings must be merged together while creating the final output for 3D model of the building. All the three tree clusters
are identified using the clustering algorithm. However, only the partial tree clusters are identified. Buildings which are classi-fied under multiple clusters are marked using MC1, MC2, and MC3. In total, all 25 building clusters including the mul-tiple building clusters are identified in dataset 1.
For dataset 2, 24 clusters were identified. The minimum and maximum number of clusters set chosen is 200 and 1200. For this dataset, there are no multiple clusters or the presence of tree clusters. Manual inspection indicates the presence of 28 buildings in dataset 2 of which 23 building clusters are identi-fied (Fig. 5).Table 3summarizes the results of Euclidean based clustering on dataset 1 and dataset 2. The choice of minimum and maximum number of points in the cluster is to minimize the over-segmentation and under-segmentation of the cluster.
4.1. Comparison with the buildings identified using Terrasolid software
For dataset 1, out of the total 33,703 points, 22,243 points were classified as non-ground points. For dataset 2, out of the total 108,198 points, 69,992 points were classified as ground points (Fig. 4). Building detection algorithm in the Terrasolid software is based on the fitting plane to a set of points which are initially identified based on the holes created identified in the ground points as discussed in Section2.5. User defined parameters such as the smallest size of the building footprint, tolerance value i.e., the minimum elevation differ-ence of a point from the plane fitted, determine the building points. In this study, the parameters were set based on heuris-tic method. The results of building modelling using Terrasolid are given inFig. 6.
A comparison of the results of the buildings modelled using Euclidean clustering based segmentation algorithm in the open source library PCL and Terrascan software algorithm is given inTables 4 and 5. As evident fromTables 4 and 5, the perfor-mance of building extraction is comparable. In dataset 1, all the 25 buildings are identified by the Euclidean clustering based segmentation algorithm whereas the commercial soft-ware is able to identify only 19 buildings. This can be attribu-ted to the failure in the detection of planar patches amongst the detected building points. In dataset 2, the commercial soft-ware outperformed the Euclidean clustering based segmenta-tion in identifying the buildings. Five building clusters marked as ‘U’ are not identified as building points. The miss-ing buildmiss-ing patches are those which have less number of Figure 5 Result of Euclidean based segmentation: (a) top view
of building clusters for dataset 2 (b) building clusters for dataset 2.
Table 3 Results of Euclidean based segmentation.
Parameters LiDAR dataset Dataset 1 Dataset 2 Cluster tolerance (dth) 1 m 0.75 m Min/max number of points in the cluster 50–5000 200–1200 Non-ground point 11,058 35,720 No. of clusters 35 24 No. of buildings in the survey area 25 25 No. of buildings identified in the survey
area
25 24
points than the minimum threshold value in the Euclidean clustering based segmentation. The minimum cluster value was set to avoid over segmentation of the building objects. From the experiments, it is evident that the choice of the user defined parameters influences substantially the building clus-ters detected. In spite of these limitations, it is evident that the building clusters detected using the proposed histogram based method in open source implementation is comparable with the commercial software. This observation signifies the functional utility of the open source point cloud library to effectively implement building detection and modelling using LiDAR point data, besides affordability. More studies are rec-ommended for automatically determining the user defined parameters based on point density of the point cloud for improving the building detection.
5. Conclusions
This study presents a novel data driven methodology for build-ing detection usbuild-ing the open source point cloud library. The point cloud data were segmented in 3D using Euclidean dis-tance based segmentation methods. From the segments created the building clusters are separated from the non-building clus-ters by a novel histogram based method. The performance of the proposed methodology has been compared with the auto-matic building detection routine available in a popular com-mercial LiDAR data processing software (Terrasolid). The performance of the proposed methodology is comparable to that of the commercial version. Overall their accuracies are 100%, 82% and 86.3%, 89.2% respectively from the two
different implementations. Further, it has been observed that the performance of the proposed methodology does not depend on the point cloud density. A major limitation of the study is that analyst needs to specify the minimum and maxi-mum number of points in a cluster. This makes the procedure a semi automatic process. Further works are recommended to refine this method to automatically select the minimum and maximum number of points in the clusters based on the dataset.
Conflict of interest
There is no conflict of interest.
Acknowledgements
The authors also acknowledge Terrasolid Ltd. and NOAA Coastal Services Center for providing the LiDAR data used in this work.
References
Anguelov, Dragomir, Dulong, Carole, Filip, Daniel, Frueh, Chris-tian, Lafon, Ste´phane, Lyon, Richard, Ogale, Abhijit, Vincent, Luc, Weaver, Josh, 2010. Google street view: capturing the world at street level. Computer 43 (6), 32–38.
Axellson, P., 2000. DEM generation from laser scanner data using adaptive TIN models. In: International Archives of the Pho-togrammetry, Remote Sensing and Spatial Information Sciences, XXXIII, pp. 110–117.
Baltsavias, E.P., 1999. Airborne laser scanning: basic relations and formulas. ISPRS J. Photogramm. Remote Sens. 54 (2–3), 199–214.
Biljecki, Filip, Stoter, Jantien, Ledoux, Hugo, Zlatanova, Sisi, C¸o¨ltekin, Arzu, 2015. Applications of 3D city models: state of the art review. ISPRS Int. J. Geo-Inf. 4 (4), 2842–2889.
Burtch, Robert, 2002. Lidar principles and applications. In: IMAGIN Conference, pp. 1–13.
Dadras, Mohsen, Shafri, Helmi Z.M., Ahmad, Noordin, Pradhan, Biswajeet, Safarpour, Sahabeh, 2015. Spatio-temporal analysis of urban growth from remote sensing data in Bandar Abbas City, Iran. Egypt. J. Remote Sens. Space Sci. 18 (1), 35–52.
Filin, S., Pfeifer, N., 2006. Segmentation of airborne laser scanning data using slope adaptive neighbourhood. ISPRS J. Photogramm. Remote Sens.
Ghosh, S., Lohani, B., 2011. Heuristical feature extraction from LiDAR data and their visualization. In: Proceedings of the ISPRS Workshop on Laser Scanning.
Lafarge, Florent, Malllet, Clement, 2012. Creating large-scale city models from 3D-point clouds: a robust approach with hybrid representation. Int. J. Comput. Vision 1, 69–85.
Lari, Z., Habib, A., Kwak, E., 2011. An adaptive approach for segmentation of 3D laser point cloud. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, pp. 103–108 (XXXVIII-5).
Meng, X., Currit, N., Zhao, K., 2010. Ground filtering algorithms for airborne LiDAR data: a review of critical issues. Remote Sens. Pfeifer, N., 2005. A subdivision algorithm for smooth 3D terrain
models. ISPRS J. Photogramm. Remote Sens.
Rabbani, T.F., Heuvel, V.D., Vosselmann, G., 2006. Egmentation of point clouds using smoothness constraints. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, pp. 248–253.
Ramiya, Anandakumar M., Nidamanuri, Rama Rao, Krishnan, Ramakrishnan, 2016. Object-oriented semantic labelling of spec-Table 4 Comparison of the buildings extracted by the
proposed EC segmentation based methodology with the Terrasolid software based results for dataset 1.
Parameters By EC segmentation
Terrasolid
Non-ground points 11,016 11,460 Building points 6433 5191 No of buildings in the surveyed area 25 25 No. of buildings identified 25 19 No. of buildings modelled 24 25 False positives 0 0 True negatives 0 0 Modelling accuracy 100 (%) 86.3 (%)
Table 5 Comparison of the buildings extracted by the proposed EC segmentation based methodology with the Terrasolid software based results for dataset 2.
Parameters By EC segmentation
Terrasolid
Non-ground points 35,720 38,206 Building points 13,676 15,966 No of buildings in the surveyed area 28 28 No. of buildings identified 23 25 No. of buildings modelled 23 20 False positives 0 0 True negatives 5 5 Modelling accuracy 82 (%) 89.2 (%)
tral–spatial LiDAR point cloud for urban land cover classification and buildings detection. Geocarto Int. 31 (2).
Rusu, R.B., 2009. Semantic 3D Object Maps for Everyday Manipu-lation in Human Living Environments. Computer Science Dept, Technische Universit, Munchen, Germany.
Rusu, Radu Bogdan, Cousins, Steve, 2011. 3D is here: Point Cloud Library (PCL). In: 2011 IEEE International Conference on Robotics and Automation (May), pp. 1–4.
Soininen Arttu, 2015. TerraScan User’s Guide.
Sun, Shaosui, Salvaggio, Carl, 2013. Aerial 3D building detection and modeling from airborne LiDAR point clouds. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.
Tner, Wolfgang Fo¨rs, 1999. 3D-city models: automatic and semiau-tomatic acquisition methods. Computing, 291–303.
Vosselman, P., 2000. Slope based filtering of laser altimetric data. In: International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXIII.
Wang, M., Tseng, Y.H., 2011a. Incremental segmentation of lidar point clouds with an octree-structured voxel space. Photogramm. Rec.
Wang, Miao, Yi-Hsing, Tseng, 2011b. Incremental segmentation of lidar point clouds with an octree-structured voxel space. Pho-togramm. Rec. 26 (133), 32–57 (March 16).
Zhang, J.X., Lin, X.G., 2012. Object-based classification of urban airborne lidar point clouds with multiple echoes using SVM. ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Informa-tion Sciences I (September), 135–140.
